# Can you list every AI agent running in your company right now? Not the approved ones. The running ones. Most companies cannot answer, and the fix is not another policy document: it is an AI agent inventory that reconciles what the company bought against what people actually run. *By Manouk Draisma ยท August 14, 2026* Canonical: https://langwatch.ai/blog/can-you-list-every-ai-agent-running-in-your-company ![Can you list every AI agent running in your company right now?](https://langwatch.ai/blog/agent-inventory-hero.webp) Not the approved ones. The running ones. Most people can't, and it's worth sitting with why. The approved list lives in a slide somewhere: the two agents procurement signed off on, the platform the CISO blessed. The running list is a different thing entirely, and nobody owns it. Here's what the running list actually looks like at most companies we talk to. A couple of agents someone built on Copilot Studio. Claude Cowork rolling out across the org because a VP liked it. A Databricks Genie that a data team stood up for one report and never turned off. Every developer on [Claude Code, Codex, or Cursor](https://langwatch.ai/blog/claude-code-usage-tracking), each on their own plan. And underneath all of that, AI quietly switching itself on inside the SaaS you already pay for: Jira, Linear, Notion, all shipping agent features into tools your people already have open. Some of this started as somebody's initiative. Most of it just became available, and no one decided anything. That's the gap between the two lists, and it grows every week without a single person approving it. ## Why policy documents don't fix this Most governance programs start by writing a policy. An acceptable-use doc, an approved-tools list, a sign-off workflow for new AI. All reasonable, and all built on an assumption that turns out to be false: that you know what's running. You can't govern what you can't list. A policy that says "only approved agents may access customer data" does nothing about the Genie a data team spun up last month, because nobody knows it exists to check it against the policy. The document describes the world you wish you had. Governance has to start from the world you actually have, which means the first artifact isn't a policy. It's an AI agent inventory. ## What an AI agent inventory has to answer An inventory of AI agents is not a list of tool names. A tool name tells you almost nothing you can act on. For each agent that's actually running, you need to answer five questions: - **Which tool, and where.** The specific agent and the platform it runs on, whether that's Copilot Studio, a Databricks workspace, or a feature inside Jira. - **Who owns it.** A person, not a team. Someone who can answer for what it does and switch it off if it misbehaves. - **What it costs.** Real spend, not the license price. A \$6,000 Claude Code seat allocation where a third of the spend is idle or side projects is a different fact than a \$6,000 bill. - **Whether anyone uses it.** Seats bought against seats active. Idle licences are the cheapest thing to cut and the easiest to miss. - **What it was for.** The initiative behind it. An agent nobody can attach a reason to is either shadow AI or a renewal waiting to be cancelled. The reason this is hard is that the answers live in different places. Licenses and subscriptions sit in procurement and finance. Actual usage sits in the tools themselves, or in nobody's hands at all. The owner is often tribal knowledge. Reconciling the two sides, what the company bought against what people actually run, is the work, and it's the part a policy document skips entirely. ## What LangWatch Governance does [LangWatch Governance](https://langwatch.ai/docs/ai-governance/overview) builds that AI agent inventory and keeps it current. It reconciles what the company bought, every license and subscription, against what people actually use, then attaches the owner and the initiative behind each agent. One catalog: the tool, the seats bought versus active, the spend, the idle waste, the owner, and the department it sits in. ![The inventory: tool chips for Claude Code, Copilot Studio, ChatGPT, Databricks Genie and more converging into one catalog with seats, licence cost, idle spend, usage, and owner per tool](https://langwatch.ai/blog/agent-inventory-graphic.webp) The point isn't the list for its own sake. It's that once the running agents are on one page with cost and ownership next to each, the rest of governance has something to stand on. You can see which agents cost the most, which nobody uses, which have no owner, and which showed up without anyone approving them. Policy comes after that, because now it applies to the world you actually have. Every figure carries where it came from and how fresh it is, because an inventory people don't trust is one they stop looking at. So, the question again, and it's a real one: can you list every AI agent running in your company right now? If you've been working on this problem for your own company, I'd like to hear how you're approaching it. We wrote down what enterprise leaders told us they need from agent governance, the six requirements and a 12-question self-assessment, in the free [Agent Governance Playbook](https://langwatch.ai/agent-governance-playbook). And if you want to see your own running list in one pane, [bring two teams' traffic to a working session](https://langwatch.ai/get-a-demo). ## Frequently asked questions ### What is an AI agent inventory? An AI agent inventory is a living catalog of every AI agent and tool actually running in an organization: which tool and where it runs, who owns it, what it really costs, whether anyone uses it, and the initiative behind it. It reconciles what the company bought (licenses and subscriptions) against what people actually use. ### Why don't AI policy documents work on their own? Because policies apply to a world you can see. A rule like "only approved agents may access customer data" does nothing about an agent nobody knows exists. You cannot govern what you cannot list, so the first governance artifact is the inventory, and policy comes after. ### What should an AI agent inventory include? Five answers per agent: the specific tool and platform it runs on, a named owner (a person, not a team), real spend rather than the license price, seats bought against seats active, and the initiative behind it. A tool name alone tells you almost nothing you can act on. ### What is shadow AI? AI tools and agents running in an organization without anyone approving or tracking them: a data-team experiment that never got turned off, agent features switching themselves on inside SaaS like Jira, Linear, and Notion, or developers on personal-plan coding agents. An agent nobody can attach an owner or a reason to is either shadow AI or a renewal waiting to be cancelled. ### How do you find idle AI spend? Compare seats bought against seats active per tool, priced. A $6,000 monthly seat allocation where a third sits idle is a different fact than a $6,000 bill, and idle licences are the cheapest thing to cut and the easiest to miss without an inventory that reconciles both sides. ### How does LangWatch Governance build the inventory? LangWatch Governance reconciles every license and subscription against actual usage, then attaches the owner and the initiative to each agent: one catalog with the tool, seats bought versus active, spend, idle waste, owner, and department. Every figure carries where it came from and how fresh it is.